senn-techKI-Radar: DeepSeek — From R1 to the V4 Series
DeepSeek has shaken up the market: first with V3, then with the reasoning-focused R1, both with open weights and impressive price-performance. The current V4 series continues this trend.
What makes DeepSeek interesting
- Open weights (for V3/R1): smaller variants can be self-hosted, a real argument for data sovereignty.
- Reasoning: R1 demonstrated that strong reasoning capability doesn't have to be expensive.
- Cost: the hosted API is affordable.
Why the R1 Line Counts
The real milestone was R1: a reasoning model that exposed its step-by-step thought path and proved that strong inference quality is not necessarily bound to a closed US frontier model. The V4 series continues that path and turns it into a standard tool rather than an experiment. For the SME it means concretely: tasks like contract analysis, technical evaluation, or multi-step planning are no longer the domain of the expensive APIs: they map cleanly onto open weights, provided the hosting is right.
Architecture: MoE With a Small Active Share
Like its predecessors, DeepSeek V4 is built on mixture-of-experts: the model holds a large total parameter count but activates only a small fraction per token. The reason for the strong price-performance lies in the ratio of total to active parameters: many billions total, only a few billion active per token. Inference cost tracks the active share, not the full model size. For self-hosting it means that even though the weights are large, the compute per call stays manageable as long as you hold enough VRAM to keep the entire model loaded. In practice that is purely a planning factor: total size dictates memory demand, the active share dictates speed.
License: MIT as a Real Lever
The open weights ship under the MIT license, one of the most permissive licenses out there. Commercial use, modification, and redistribution are allowed without traps, and there is no user cap like the ones some other vendors impose. For companies that want to embed a model in a product, that is a tangible advantage over more restrictive licenses. Weights are available on Hugging Face, so getting started needs no negotiation.
The important caveat
The hosted DeepSeek API is operated from China. For personal or confidential data, this is a clear GDPR issue: inputs leave the EU. Anyone wanting to use DeepSeek in a data protection-compliant way hosts the open models themselves (vLLM, Ollama). Quantization to 4-bit or FP8 keeps memory demand in check at a small quality cost, a familiar trade-off that should be weighed per use case. The key distinction: the open model and the cloud API are two different things, and only the first is an option for sensitive data.
In our own case
The live demo on this website runs via our own LiteLLM gateway on our own hardware in the EU — no third country, no data shared with external APIs. Exactly this clean variant (self-hosting instead of foreign cloud API) is what we also recommend to our customers. DeepSeek is a well-suited candidate for that: permissively licensed, efficient as an MoE, and available in several sizes.
Further Reading
Can we use DeepSeek in a GDPR-compliant way in the EU?+
Yes, but not via the hosted API — it is operated from China, inputs leave the EU, and that is a clear GDPR issue. Data-protection-compliant use works only through self-hosted open models with vLLM or Ollama on your own hardware in the EU. Quantization to 4-bit or FP8 keeps memory demand in check at a small quality cost.
Why is DeepSeek's MIT license an advantage?+
MIT is one of the most permissive licenses out there. Commercial use, modification, and redistribution are allowed without traps, and there is no user cap as with some other vendors. For companies that want to embed a model in a product, that is a tangible advantage. The weights are on Hugging Face, so getting started needs no negotiation.
What does mixture-of-experts mean for self-hosting?+
DeepSeek V4 activates only a small fraction of its many parameters per token. Inference cost tracks the active share, not the full model size — the reason for the strong price-performance. For self-hosting that means: total size dictates memory demand (lots of VRAM), the active share dictates speed. Something to plan for, nothing more.